Fig. 4: Connectivity analysis and deep-learning-based synapse detection.

a, Spiny dendrite with SHANK2-expressing excitatory (magenta) and gephyrin-expressing inhibitory post-synapses (yellow), connected axons (presynaptic bassoon: cyan) and seed locations. b, Input density onto spiny dendrites (primary somatosensory cortex), defined by immunolabelling: excitatory (SHANK2⁺); inhibitory (gephyrin⁺); IF⁺: total immunofluorescence. Numerical values: mean ± s.d. throughout; box plots: median; lower and upper quartiles; whiskers: minimum and maximum (throughout). Data points: individual dendrites (11; 123 µm total, 351 IF⁺ synapses, b–f). c, Synapse (SHANK2⁺ or gephyrin⁺) target locations. d, Molecular properties of spine heads, with synapses identified by immunolabelling or PSDs. e, Molecular properties of shaft synapses. f, Excitatory (E) versus inhibitory (I) synapses onto dendrites. g, Spine seeding by excitatory axon outputs. h, Output density of spine-seeded axons (19 axons (data points) with two or more outputs, total length 990 µm). i, Spine fraction traced to parent dendrite for spines contacted by the same 19 axons (128 spines, mean ± s.d. over individual axon). Data points: axons. j, Inhibitory axons seeded at an AIS (gephyrin⁺, no PSD). k, Output fraction onto shafts of AIS-seeded axons (eight axons with three or more outputs). l, Deep-learning prediction of bassoon and SHANK2 location from structural LICONN channel. Single plane from LICONN volume (CA1, stratum radiatum, not included in training), comparing prediction with immunolabelling. Scale bars, 1 μm. m, Corresponding volumetric renderings. n, Excitatory input and output for a pyramidal neuron from the dataset in Fig. 1a. Synapse detections through bassoon and SHANK2 prediction mapped onto FFN segmentation. Magenta numbers: detected synapses between indicated branch points. Magnified views: structural channel, molecule predictions and cellular segments (partial proofreading, eliminating false-positive detections without adding missed detections). Scale bars, 2 μm (top); 20 μm (middle); 500 nm (bottom). o, Connectivity prediction in hippocampal dataset from Fig. 2a. Left, rendered synapse predictions. Middle, axon (black) with connected dendrites. Right, dendrite (black) with connected axons. Scale bar, 10 μm. p, Integration of structural, immunolabelling and deep-learning analysis. Dendrite (primary somatosensory cortex) with immunolabelling-based detection of excitatory (501) and inhibitory (80) post-synapses. Insets: presynaptic partners identified by deep-learning bassoon prediction. Scale bar, 10 μm.

Source Data